OpenAI’s Astra: How an Unreleased Model Solved 10 Open Math Problems for $2,000

A Model With No Release Date Just Out-Proved Human Mathematicians

On August 1, 2026, OpenAI announced that an internal, unreleased version of its next model — code-named Astra — had solved 10 problems in mathematics and theoretical computer science that had remained open for a decade or more. The problems spanned group theory, von Neumann algebras, combinatorics, high-dimensional sphere packing, coding theory, arithmetic circuit complexity, quantum parallel repetition, and lattice cryptography, according to SiliconANGLE’s summary of the release.

The headline number is small but pointed: Astra generated the ten proofs for approximately $2,000 in compute at GPT-5.6 API rates, per SiliconANGLE’s reporting. That figure matters because it reframes the announcement from “a lab spent months on a moonshot” to “a model did this as a side task at the cost of a laptop.”

This was not Astra’s first appearance. Forbes reported that the same model family disproved the 80-year-old Erdős unit distance conjecture back in May 2026 — a result Thomas Bloom, maintainer of the Erdős problem catalogue, called “big news” at the time. August’s release of ten proofs at once turned an isolated result into a pattern.

That pattern carries a caveat worth stating up front: OpenAI has overstated Erdős-related results before. In October 2025, OpenAI VP Kevin Weil claimed GPT-5 had “found solutions to 10 (!) previously unsolved Erdős problems,” but Bloom — the same mathematician who later praised the May 2026 result — publicly disputed that framing, clarifying that problems marked “open” on his site meant only that he personally was unaware of a published solution, not that none existed. GPT-5 had found existing literature Bloom hadn’t catalogued, not solved unsolved conjectures — a distinction Meta’s Yann LeCun and Google DeepMind’s Demis Hassabis both called out publicly at the time, per TechCrunch’s reporting.

The Non-Sofic Group: A 27-Year-Old Open Question

The most cited result in the batch is the first explicit construction of evidence for a non-sofic group — a concept introduced by mathematician Mikhail Gromov in 1999, per SiliconANGLE’s reporting, making the underlying question 27 years old at the time of Astra’s result. Sofic groups are a class of infinite groups that can be approximated by finite structures in a precise technical sense; whether every group is sofic, or whether non-sofic groups exist at all, had resisted resolution by human mathematicians for nearly three decades.

The batch also touched Connes’s rigidity conjecture in von Neumann algebras, extremal problems in Ramsey theory and graph theory, and the closest vector problem in lattice cryptography — a foundational hardness assumption underlying several post-quantum encryption schemes. Each result arrived not as a plain-language claim but as a fully formalized Lean 4 proof.

Zero “Sorry” — What Machine-Checkable Actually Means

The distinction between “an AI said it solved a problem” and “an AI produced a verifiable proof” is the substance of this story. Lean is a proof assistant: a piece of software that checks, line by line, whether each logical step in a proof actually follows from the previous ones and from established axioms. A Lean proof with an unfilled step contains the placeholder keyword “sorry” — a flag that tells anyone reading the code exactly where the argument is incomplete.

According to SiliconANGLE, the “sorry” count across all ten formalized proofs is zero: no step in any of the ten has been left unproven. OpenAI published the certificate files on GitHub under an Apache 2.0 license, meaning any researcher — with no OpenAI account, no API key, and no trust required in the company’s claims — can download Lean, load the files, and watch the verifier confirm every step independently.

That auditability is what drew the strongest endorsement. Fields Medal winner Timothy Gowers said he would recommend one of the proofs for a top mathematics journal without hesitation, according to Forbes’ reporting on the release.

The Skeptics Have a Point Too

Not every response was celebratory. Cognitive scientist and AI critic Gary Marcus called the release “amazing but vastly oversold,” per Forbes, arguing that OpenAI selected which problems to publicize — success stories from a much larger, unpublicized batch of attempts — and that outside researchers cannot independently run the model to test it on new problems of their own choosing. Both points are structural rather than about the specific proofs: the Lean verification confirms the ten results are correct, but it says nothing about the failure rate on problems that did not make the announcement, and it says nothing about what Astra can do beyond math, since no one outside OpenAI has hands-on access.

That access gap is not incidental. Forbes reported that Astra has no announced release date, no pricing, and must clear a US government security review before any public rollout. Days later, OpenAI disclosed the reason for the delay was more specific than routine caution: the company said it had slowed Astra’s development after the model reached a “critical cybersecurity threshold,” meaning it could independently identify and carry out cyberattacks against well-protected real-world systems. OpenAI said it is now working with government agencies and AI safety organizations on stricter controls before any release — meaning the model that produced these math results is not something a university lab, a startup, or a rival AI company can currently test, replicate the workflow of, or benchmark against.

What This Signals for AI’s Trajectory

The practical significance is less about ten specific theorems and more about what the cost and verification method imply for where AI-assisted research is heading.

First, the price point. A $2,000 compute bill for ten decade-plus-old open problems — even a self-selected, favorable sample — is a number research institutions and even well-funded individual mathematicians can absorb. If the technique generalizes even partially, it changes the economics of attacking backlogs of open conjectures across mathematics and theoretical computer science, fields where human researcher-hours are the binding constraint.

Second, the verification method matters more than the headline. Machine-checkable proofs sidestep the core trust problem that has dogged AI-generated claims in every other domain: instead of asking readers to believe a language model’s prose explanation, Lean formalization forces every logical step to survive an independent, deterministic checker. That is a meaningfully different evidentiary standard than an AI chatbot asserting a fact, and it is why a Fields Medalist was willing to vouch for the output without having watched the model work.

Third, the gap between capability and access is likely to widen before it narrows. A security review gating a model’s public release — on top of the usual competitive incentive to keep frontier capability internal — means the tools capable of this kind of result are increasingly things labs use, not things researchers everywhere can rent by the hour. That has direct implications for which institutions can benefit first from AI-accelerated mathematics, and which are left waiting for a public API.

What This Means for Research Institutions Weighing AI-Assisted Discovery

1. Treat formal verification as the new bar for AI-generated scientific claims

Any AI-assisted result that cannot be independently verified by a deterministic tool — Lean, Coq, Isabelle, or an equivalent — should be treated as a hypothesis, not a finding. The Astra release is instructive precisely because it did not ask anyone to trust OpenAI; it asked them to run a free, open-source checker. Institutions building internal AI-research workflows should adopt the same discipline: pair any LLM-generated proof, code artifact, or dataset claim with an automated, third-party-auditable verification step before treating it as settled.

2. Budget for AI-research compute as a rounding error, not a moonshot line item

At roughly $2,000 for ten open-problem attempts, the marginal cost of pointing a frontier model at a specific open conjecture is now within reach of individual faculty research budgets, not just well-funded labs. Universities and research consortia in Algeria and across the Global South that have historically been priced out of large-scale computational research should re-examine whether targeted, narrow AI-assisted attempts on specific open problems in their researchers’ specialties are now affordable pilot projects.

3. Plan for a widening gap between frontier capability and public access

Astra’s security-review gate means the most capable systems will likely reach the public later, and with more restrictions, than earlier AI generations. Institutions should not assume that “wait six months and rent the API” is a reliable strategy for accessing frontier mathematical-reasoning tools; national research funders should instead build relationships and pilot programs directly with frontier labs where possible, rather than assuming eventual commodity access.

4. Separate the marketing narrative from the falsifiable claim

Gary Marcus’s critique — that problem selection is favorable and outside testing is impossible — is a reminder to read past the headline number. When evaluating any AI lab’s research claims, ask specifically: how many attempts failed, is the model or its outputs independently testable, and is the verification method itself trustworthy and open. The Astra case scores well on the last question and poorly on the first two, and both parts of that assessment are true at once.

The Wider Pattern: AI Research Tools Outpacing AI Research Access

The Astra proofs sit inside a broader 2026 pattern: frontier AI capability announcements are increasingly framed around narrow, verifiable, high-prestige demonstrations — a math proof, a coding benchmark, a scientific hypothesis — rather than broad consumer product launches. That shift makes sense commercially: a Fields Medalist’s endorsement is a more durable credibility signal than a chatbot demo, and it costs a fraction of a foundation model’s total training budget to generate. But it also means the public conversation about AI progress is increasingly shaped by results the public cannot reproduce, tools the public cannot rent, and models the public cannot interrogate. The Lean verification is a genuine, durable check on the specific claims made on August 1 — the ten proofs are correct, full stop. What it cannot check is everything the announcement chose not to show.

Frequently Asked Questions

Sources & Further Reading

• OpenAI’s Astra solves 10 long-open math problems and publishes the proofs — SiliconANGLE
• OpenAI’s Astra Solved 10 Decades-Old Math Problems For $2,000 — Forbes
• OpenAI says it slowed Astra model development over security concerns — TechCrunch
• OpenAI’s embarrassing math (GPT-5 Erdős problems dispute) — TechCrunch